A fault prevention and diagnosis control system based on intelligent mine shield tunneling machine of minehong

Abstract

This invention relates to the field of industrial automation technology, specifically disclosing a fault prevention and diagnosis control system for tunnel boring machines (TBMs) based on the Kuanghong intelligent mining platform. The system utilizes the Kuanghong operating system to achieve spatiotemporal synchronous acquisition and fusion of data from multiple heterogeneous sensors, constructing a multi-dimensional state sequence including equipment operating parameters and health indicators. Based on historical data, a probabilistic world model is established to predict state transition probability distributions and quantify uncertainties. Safe zones are defined according to uncertainties, and the safety of actions is verified through multi-step rolling prediction. A safety constraint strategy optimization algorithm is used to iteratively update the decision-making strategy, and online learning mechanisms are combined to dynamically update model parameters. This invention solves the problems of insufficient data fusion, lagging fault prediction, and inadequate exploration of action safety in existing technologies. It achieves accurate perception of the TBM's operating status, early warning of faults, and safe autonomous decision-making, improving equipment operating reliability and maintenance efficiency.

Description

Technical Field

[0001] This invention relates to the field of industrial automation technology, specifically to a fault prevention and diagnosis control system for mining tunnel boring machines based on Kuanghong Intelligent. Background Technology

[0002] As mining extends to deeper and more complex geological conditions, the reliability and safety of tunnel boring machines (TBMs), as key excavation equipment, are of paramount importance. Traditional TBM fault diagnosis and maintenance primarily rely on periodic inspections, threshold alarms, and manual experience. This approach suffers from significant delays, struggles to handle sudden failures, and is prone to resource waste or safety accidents due to over- or under-maintenance. In recent years, with the development of Industrial Internet of Things (IIoT) and Artificial Intelligence (AI) technologies, data-driven predictive maintenance has become a hot research topic in the industry. By deploying sensor networks to collect equipment operating data and utilizing machine learning algorithms to build fault prediction models, theoretically, early warning and diagnosis of faults can be achieved, thereby improving the intelligence level of equipment operation and maintenance.

[0003] The existing technology has the following shortcomings:

[0004] In the context of health management for mining equipment, the normalized, fault-free operation prevents the reward function from providing sufficient and effective gradient signals, driving the agent to execute exploratory strategies in search of higher rewards. This exploratory behavior in the physical world may manifest as unverified fine-tuning of system control parameters. If this micro-operation resonates with the inherent frequency of the equipment's mechanical structure, it will induce high-frequency alternating stresses that are difficult to monitor in real time, causing cumulative fatigue damage to critical components. This damage exhibits significant hysteresis and insidiousness; its failure mode deviates from the agent's learning framework based on instantaneous rewards, making the causal chain of failure difficult to trace. This creates the potential for sudden, severe failures, constituting a core obstacle to the application of existing data-driven reinforcement learning techniques in high-end equipment. Summary of the Invention

[0005] The purpose of this invention is to provide a fault prevention and diagnosis control system for mining shield machines based on Kuanghong Intelligent, so as to solve the problems mentioned above.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A fault prevention and diagnosis control system for mining shield machines based on Kuanghong Intelligent includes:

[0008] The multi-source data acquisition and fusion module collects multi-source heterogeneous sensor data in real time during the operation of the tunnel boring machine through the Kuanghong operating system, and constructs a system status sequence that includes equipment operating parameters and health indicators.

[0009] The probabilistic world model building module establishes a probabilistic world model based on the system state sequence and historical action sequence, which can predict the state transition probability distribution and output the quantified value of the prediction uncertainty.

[0010] The safe zone dynamic division module divides the state action space into a high-confidence safe zone and a low-confidence unknown zone based on the predicted uncertainty quantification value.

[0011] The action safety simulation verification module simulates the execution of candidate actions through the probabilistic world model when the agent generates candidate actions, and calculates the cumulative uncertainty on the predicted state path.

[0012] The safety action decision and execution module compares the accumulated uncertainty with a preset safety threshold: if it is lower than the threshold, the candidate action is output to the control system as a safe and executable action; if it exceeds the threshold, the candidate action is blocked and a preset safety strategy is triggered to generate an alternative action.

[0013] The online learning and optimization module feeds back the actual safety actions performed and the resulting state transition data to the probabilistic world model, dynamically updating the model parameters and expanding the boundaries of the safety region.

[0014] The intelligent decision-making strategy optimization module uses a strategy optimization algorithm to iteratively update the agent's decision-making strategy based on the state transition results after the execution of safety actions, thereby continuously optimizing the fault prevention and autonomous diagnosis functions.

[0015] As a further aspect of the present invention: the process of obtaining the system state sequence is as follows:

[0016] Using the soft bus protocol of the Kuanghong operating system, the raw measurement values ​​of vibration, temperature, pressure and displacement sensors deployed in the main drive system, propulsion system, cutterhead system, hydraulic system and lubrication system of the tunnel boring machine are collected in parallel, and the data of each system are marked with a unified time stamp.

[0017] The raw measurement values ​​are preprocessed in real time to eliminate measurement noise caused by downhole electromagnetic interference and mechanical shock. At the same time, the effective value, peak factor and kurtosis index of each sensor are calculated by feature extraction to form a set of equipment health indicators.

[0018] The preprocessed sensor measurements and the set of equipment health indicators are combined in time sequence to construct a multi-dimensional state vector containing instantaneous operating parameters and medium- to long-term health trends. Each dimension of the state vector is normalized.

[0019] The multidimensional state vector is continuously updated with a fixed sampling period to form a system state sequence that can reflect the complete operating status of the device in real time, and the system state sequence is transmitted to the data storage center.

[0020] As a further aspect of the present invention: the construction process of the probabilistic world model is as follows:

[0021] The multidimensional state vectors in the system state sequence are combined with the historically executed action sequences according to the time correspondence to form a state-action pair sequence, which is used as the model training input.

[0022] A state transition probability model is established using a non-parametric method based on kernel density estimation. By calculating the similarity weights of state-action pairs in the feature space, the probability distribution of the next state under a given state-action condition is predicted.

[0023] The uncertainty of the probabilistic world model for each state-action pair is quantified by calculating the degree of dispersion of adjacent sample points in the predicted state distribution. The quantified uncertainty value is represented by a normalized value between zero and one.

[0024] The probabilistic world model is updated online using state transition data within a recent time period, enabling it to adapt to changes in device state over time.

[0025] As a further aspect of the present invention: the division of the state-action space into a high-confidence safe region and a low-confidence unknown region specifically includes:

[0026] Determine whether the quantified value of the prediction uncertainty is greater than or equal to a preset threshold. If yes, it is recorded as a high-confidence safe zone; otherwise, it is recorded as a low-confidence unknown zone.

[0027] The preset threshold is determined through historical data statistics or experimental calibration, and its specific value is in the range of 0.6 to 0.8, preferably 0.7.

[0028] As a further aspect of the present invention: the step of simulating the execution of candidate actions through the probabilistic world model and calculating the cumulative uncertainty on the predicted state path specifically includes:

[0029] Starting from the current state, the candidate actions are input into the probabilistic world model to obtain the next state prediction, and the next state prediction is used as a new starting point to iteratively execute multi-step state transition prediction, forming a multi-step predicted state path.

[0030] In each prediction step, the state prediction uncertainty quantification value is recorded, and the cumulative uncertainty on the entire prediction path is calculated using a time decay weighted algorithm. The uncertainty of the recent prediction steps is assigned a weight of 0.9 to 1.0, and the uncertainty of the long-term prediction steps is assigned a weight of 0.1 to 0.4.

[0031] When a critical state node is predicted, multiple possible action branches are generated to continue prediction. The cumulative uncertainty of each branch path is calculated and the maximum value is taken as the final cumulative uncertainty of the corresponding candidate action.

[0032] The multi-step prediction path and the uncertainty quantification values ​​of each node are displayed in the form of time-series animation, providing an intuitive security verification process.

[0033] As a further aspect of the present invention, the optimization and adjustment function for the candidate action specifically includes:

[0034] When the accumulated uncertainty of the candidate action exceeds the safety threshold, the action fine-tuning algorithm is activated to generate multiple alternative actions with small variations in the neighborhood of the candidate action.

[0035] Each alternative action is sequentially input into the probabilistic world model for multi-step prediction, and the cumulative uncertainty corresponding to each alternative action is calculated.

[0036] The alternative action with the lowest cumulative uncertainty and below the safety threshold is selected as the optimized safety action. If the cumulative uncertainty of all alternative actions exceeds the threshold, the preset safety policy is activated.

[0037] Record the process data for each action optimization and adjustment, including the original action, alternative actions, and their uncertainties, to form an action optimization knowledge base.

[0038] As a further aspect of the present invention: the step of feeding back the actual security actions and the resulting state transition data to the probabilistic world model, dynamically updating the model parameters and expanding the security region boundary, specifically includes:

[0039] By comparing the actual state transitions with the model's predicted states, the state transition data of the top 20% percentiles with prediction errors greater than all sample error values ​​are selected as high-value samples.

[0040] The high-value samples obtained from the screening are input into the probabilistic world model as streaming data, and only the local parameters of the model associated with these samples are updated in a targeted manner.

[0041] When new data conflicts with existing model knowledge, the model structure adaptive adjustment process is initiated, and new kernel functions are added to adapt to changes in state transition rules.

[0042] The prediction uncertainty of all known state-action pairs is recalculated based on the updated probabilistic world model, and the boundary division between the safe zone and the unknown zone is dynamically adjusted.

[0043] As a further aspect of the present invention: the continuous optimization of the fault prevention and autonomous diagnosis functions specifically includes:

[0044] Based on the degree of improvement in equipment health indicators and the extent of reduction in failure risk in the state transition results, different priority weights are assigned to each state transition sample.

[0045] When updating policy network parameters, introduce security constraints to ensure that the probability of the new policy's action selection within the safe zone is not lower than a set threshold.

[0046] Implement periodic policy distillation and transfer to distill the knowledge of a well-trained master policy network into a lightweight policy network, enabling rapid deployment and real-time decision-making;

[0047] Regularly test the effectiveness of new policies in real-world environments, and automatically roll back to the previous stable version when performance degradation or security risks are detected.

[0048] The beneficial effects of this invention are:

[0049] (1) This invention utilizes the soft bus protocol of the Kuanghong operating system to achieve spatiotemporal synchronous acquisition and integration of multi-source heterogeneous sensor data. Adaptive filtering and feature extraction algorithms are used to denoise the raw data, and a state vector containing 128 dimensions is constructed, which can simultaneously characterize the instantaneous operating parameters and medium- and long-term health trends of the equipment. This processing method improves the accuracy and comprehensiveness of the perception of the equipment's operating status, providing a more reliable data foundation for subsequent fault prediction.

[0050] (2) By predicting the uncertainty of state transitions through a probabilistic world model, a dynamic partitioning mechanism for safe regions is established, and safety constraints are introduced during the policy optimization process to ensure that the agent makes action selections within the safe regions. At the same time, action safety simulation verification and multi-branch prediction evaluation are adopted to effectively avoid the execution of high-risk actions and improve the decision-making safety and reliability of the system under complex working conditions. Attached Figure Description

[0051] The invention will now be further described with reference to the accompanying drawings.

[0052] Figure 1 This is a flowchart of the system of the present invention. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] Please see Figure 1 As shown, this invention is a fault prevention and diagnosis control system for mine tunnel boring machines based on Kuanghong Intelligent, comprising:

[0055] The multi-source data acquisition and fusion module collects multi-source heterogeneous sensor data in real time during the operation of the tunnel boring machine through the Kuanghong operating system, and constructs a system status sequence that includes equipment operating parameters and health indicators.

[0056] The probabilistic world model building module establishes a probabilistic world model based on the system state sequence and historical action sequence, which can predict the state transition probability distribution and output the quantified value of the prediction uncertainty.

[0057] The safe zone dynamic division module divides the state action space into a high-confidence safe zone and a low-confidence unknown zone based on the predicted uncertainty quantification value.

[0058] The action safety simulation verification module simulates the execution of candidate actions through the probabilistic world model when the agent generates candidate actions, and calculates the cumulative uncertainty on the predicted state path.

[0059] The safety action decision and execution module compares the accumulated uncertainty with a preset safety threshold: if it is lower than the threshold, the candidate action is output to the control system as a safe and executable action; if it exceeds the threshold, the candidate action is blocked and a preset safety strategy is triggered to generate an alternative action.

[0060] The online learning and optimization module feeds back the actual safety actions performed and the resulting state transition data to the probabilistic world model, dynamically updating the model parameters and expanding the boundaries of the safety region.

[0061] The intelligent decision-making strategy optimization module uses a strategy optimization algorithm to iteratively update the agent's decision-making strategy based on the state transition results after the execution of safety actions, thereby continuously optimizing the fault prevention and autonomous diagnosis functions.

[0062] The autonomous fault diagnosis identifies potential fault risks by analyzing abnormal patterns and predictive uncertainties in the system state sequence, and blocks or mitigates fault development through safe action decisions.

[0063] In the multi-source data acquisition and fusion module, real-time acquisition and integration of multi-source heterogeneous sensor data are achieved through the soft bus protocol of the Kuanghong operating system. Specifically, this module establishes communication connections with sensors of various subsystems of the tunnel boring machine through the distributed data bus service provided by the Kuanghong operating system. These sensors include vibration acceleration and temperature sensors deployed in the main drive system, pressure and displacement sensors in the propulsion system, torque and speed sensors in the cutterhead system, oil pressure and flow sensors in the hydraulic system, and oil quality and level sensors in the lubrication system. During data acquisition, the module reads the raw measurement values ​​of each sensor in parallel at a period of 10 milliseconds and adds a unified timestamp to each data packet through the global clock service of the Kuanghong operating system to ensure the time synchronization of multi-source data. The acquired raw data includes the time-domain waveform of the vibration signal, the temperature sensor reading in degrees Celsius, the pressure sensor reading in kilopascals, the displacement sensor reading in millimeters, and other physical quantity measurement results.

[0064] In the data preprocessing stage, this module employs digital signal processing (DSP) technology to denoise and extract features from the raw measurements. For vibration signals, a sliding window-based amplitude limiting filter algorithm is first used to eliminate abnormal pulse interference, followed by a fourth-order Butterworth low-pass filter to remove high-frequency noise, with the cutoff frequency set to 1.2 times the highest effective frequency of the signal. For slowly varying signals such as temperature and pressure, a median filtering algorithm is used to eliminate occasional measurement errors. Based on signal preprocessing, the module calculates the characteristic indicators of various sensors in real time: for vibration signals, the root mean square value within a 100-millisecond time window is calculated as the effective value, the ratio of the peak value to the effective value is calculated as the peak factor, and the ratio of the fourth-order central moment to the fourth power of the standard deviation is calculated as the kurtosis index; for signals such as temperature and pressure, the arithmetic mean and standard deviation within a 1-second time window are calculated. These characteristic indicators collectively constitute the equipment health indicator set, used to characterize the operating status and degree of degradation of the equipment.

[0065] In the state vector construction phase, this module performs spatiotemporal fusion of preprocessed sensor measurements and equipment health indicator sets. First, various data points from the same moment are aligned using a unified timestamp, forming data frames containing multiple dimensions. Each data frame includes instantaneous operating parameters such as raw measurements from vibration sensors, temperature sensors, and pressure sensors, as well as health indicators such as effective vibration values, peak factor, and kurtosis. Subsequently, the data for each dimension is normalized using a maximum-minimum normalization method to convert the values ​​to between 0 and 1, eliminating the influence of dimensions. The final constructed multidimensional state vector contains 128 dimensions, with 64 dimensions representing raw sensor measurements and 64 dimensions representing equipment health indicators, comprehensively characterizing the equipment's instantaneous operating status and medium- to long-term health trends.

[0066] During the system state sequence generation phase, this module continuously updates the multidimensional state vector at a fixed sampling period. The system uses a 100-millisecond sampling period, generating a new state vector each period and storing it sequentially in a circular buffer. Each state vector carries a precise timestamp, forming continuous time-series data. This sequence reflects complete information about the device's operating status in real time, including normal operation, abnormal states, and transitional states. The final generated system state sequence is transmitted to the data storage center via the distributed data service of the Mining Hong operating system for use by the subsequent probabilistic world model construction module. The data storage center uses a time-series database for data storage, supporting high-speed write and query operations to ensure the integrity and availability of the system state sequence.

[0067] The entire data acquisition and fusion process adopts a pipelined architecture, with each processing stage executing in parallel to ensure that continuously generated sensor data can be processed in real time. The module also features an anomaly handling mechanism; when a sensor malfunction or communication interruption is detected, a data reconstruction algorithm is automatically activated to estimate the state using data from adjacent sensors, ensuring the continuity of the system state sequence. Simultaneously, the module provides data quality monitoring functionality, calculating the signal-to-noise ratio and availability metrics for each data channel in real time, providing data quality references for subsequent analysis and processing.

[0068] In the probabilistic world model construction module, the mathematical model capable of predicting system state transition patterns and quantifying prediction uncertainties is established. This module first preprocesses the training data, aligning the system state sequence generated by the multi-source data acquisition and fusion module with the recorded historical action sequences in a spatiotemporal manner. Specifically, the module obtains a multi-dimensional state vector sequence with timestamps from the data storage center and extracts execution action records for the corresponding time periods from the control log, including control commands such as feed pressure adjustment, cutterhead speed adjustment, and hydraulic system parameter settings. Through precise timestamp matching, the state vector at each moment is combined with the executed action to form a state-action pair, creating a sequence of state-action pairs arranged chronologically. To ensure data quality, the module automatically detects and removes time periods with missing or abnormal data, ensuring the integrity and reliability of the training data.

[0069] The core technology of this module is to establish a state transition probability model using a non-parametric method based on kernel density estimation. During model construction, historical state-action pairs are first mapped to a high-dimensional feature space. The state transition probability distribution is estimated by calculating the similarity between sample points in the feature space. For a given state-action pair, the model searches for its most similar nearest neighbors in historical data. Based on the results of these nearest neighbors after the state transition, the probability distribution of the next state is calculated using a weighted kernel function. The kernel function uses a Gaussian kernel, and its bandwidth parameter is determined through cross-validation to ensure accurate estimation of the state transition probability distribution. The advantage of this non-parametric method is that it does not require prior assumptions about the state transition rules and can adaptively capture complex nonlinear relationships.

[0070] The quantification of prediction uncertainty is achieved by analyzing the characteristics of the estimated probability distribution. For each state-action pair, the model calculates the dispersion of sample points in its predicted state distribution, specifically by calculating the distribution entropy value to measure uncertainty. First, a large number of sample points are randomly generated in the predicted state distribution. Then, the variance of these sample points in each dimension is calculated. Finally, the variance values ​​of each dimension are weighted and combined into a normalized uncertainty index between 0 and 1. The closer the value is to 0, the more certain the model's prediction of the state transition under that state-action is; the closer the value is to 1, the higher the prediction uncertainty. This quantified uncertainty value will serve as an important basis for subsequent safety decisions, helping the system identify high-risk state-action regions.

[0071] To enable the probabilistic world model to adapt to changes in device state over time, the module employs a sliding time window online update mechanism. The system maintains a data window of fixed length, such as state transition data for the most recent 30 days. When new state transition data is generated, it is automatically added to the training dataset, while older data exceeding the time window is removed. The model update process uses incremental learning, automatically triggering model parameter updates at regular intervals (e.g., 24 hours) or when a certain amount of new data has accumulated. During the update process, the model focuses on learning the patterns of device state changes reflected in the new data, while retaining general knowledge from historical data. This ensures that the model can both track the latest changes in device state and retain important historical experience.

[0072] This module also provides model performance monitoring and validation capabilities. It periodically uses a portion of the latest data as a test set to evaluate the model's predictive accuracy. Monitoring metrics include the average error of state predictions and the degree of calibration of uncertainty estimates. When a decline in model performance is detected, it automatically adjusts model parameters or triggers a retraining process. Simultaneously, the module records historical model versions and corresponding performance metrics, supporting model version retrospection and comparison, providing the system with reliable model management capabilities. The entire model building and update process employs an automated pipeline design, requiring no manual intervention, ensuring that the probabilistic world model continuously maintains accurate predictive capabilities.

[0073] It should be noted that the preset safety threshold is determined through statistical analysis of historical fault data and calibration based on expert experience, and is specifically set to 0.7. This threshold represents the system's confidence level in predicting state transitions. When the uncertainty quantification value is below 0.7, the prediction is considered reliable and the action is safe; otherwise, it is considered high-risk.

[0074] In the dynamic division module of safe areas, it is determined whether the quantified value of prediction uncertainty is greater than or equal to the preset threshold (set to 0.7). If so, it is recorded as a high-confidence safe area; otherwise, it is recorded as a low-confidence unknown area.

[0075] The preset threshold is determined through historical data statistics or experimental calibration.

[0076] In the action safety simulation and verification module, the safety assessment of candidate actions generated by the agent is performed. It simulates the action execution process using a probabilistic world model and calculates the accumulated uncertainty along the prediction path. Upon startup, the module first acquires the multi-dimensional state vector of the current device state and the candidate actions to be evaluated. Using the current state as the initial point, the candidate actions are input into the probabilistic world model for state transition prediction, resulting in a state prediction for the next time step. This prediction result serves as the new starting state, and the same candidate actions are input again, or subsequent actions are generated according to a preset strategy. This iterative process of multiple state transition predictions is performed, forming a predicted state path with a length of 20 time steps. Throughout the prediction process, the module records the predicted state vector and its corresponding timestamp for each time step, constructing a complete predicted state trajectory.

[0077] During each state prediction step, the module synchronously records the uncertainty quantification value output by the probabilistic world model. To assess the overall risk level along the entire prediction path, a time-decay weighted algorithm is used to calculate the cumulative uncertainty. This algorithm assigns different weight coefficients to uncertainties at different time steps along the prediction path: higher weights (0.9 to 1.0) are given to recent prediction steps (e.g., the first 5 time steps), medium weights (0.5 to 0.8) are given to intermediate prediction steps (6th to 15th time steps), and lower weights (0.1 to 0.4) are given to long-term prediction steps (16th to 20th time steps). In the weighted calculation, the uncertainty quantification value at each time step is multiplied by its corresponding weight coefficient, summed, and then divided by the sum of all weight coefficients to obtain a normalized cumulative uncertainty index between 0 and 1. This weighting method reflects the system's emphasis on the reliability of recent prediction results and meets the actual safety requirements of the control system.

[0078] When a critical state node is encountered during the prediction process, the module initiates a branch expansion mechanism. Critical state nodes include situations where equipment operating parameters are approaching safety thresholds, health indicators show abnormal trends, or the device is in a known high-risk area. Upon detecting a critical state node, the module generates 3 to 5 possible subsequent action branches. These branches include different options such as continuing the original action, reducing the magnitude of the action, or switching to a safety strategy. For each action branch, multi-step state prediction is performed, and the cumulative uncertainty of each branch path is calculated separately. Finally, the maximum cumulative uncertainty across all branch paths is taken as the final risk assessment result for that candidate action. This branch expansion mechanism ensures the conservatism and comprehensiveness of the safety assessment, avoiding potential risks that might be overlooked by a single prediction path.

[0079] To provide an intuitive safety verification process, the module integrates visualization functionality. This function presents the multi-step prediction path and the quantified uncertainty values ​​of each node in a time-series animation, using different colors and transparency to visually display the risk level at each time step. State variables in the prediction path are displayed as line graphs, with a color gradient representing the magnitude of uncertainty, transitioning from green (low uncertainty) to red (high uncertainty). Users can view detailed state information and uncertainty values ​​at any time step through an interactive interface, and can also adjust parameters such as the prediction step size and the number of branches. The visualization system also supports playback and comparison of historical prediction records, facilitating analysts' understanding of the system's decision-making process and safety assessment results. All visualization data is updated in real time, keeping pace with the simulation verification process.

[0080] It should be noted that when calculating cumulative uncertainty, the uncertainty of the near-term forecast step (the first 5 time steps) is given a higher weight of 0.9 to 1.0, while the uncertainty of the far-term forecast step (the 16th to 20th time steps) is given a lower weight of 0.1 to 0.4, in order to reflect the importance attached to the near-term forecast results.

[0081] In the safety action decision and execution module, the accumulated uncertainty is compared with a preset safety threshold: if it is lower than the threshold, the candidate action is output to the control system as a safe and executable action; if it exceeds the threshold, the candidate action is blocked and a preset safety strategy is triggered to generate an alternative action.

[0082] In the online model learning and optimization module, the continuous improvement and optimization of the probabilistic world model is achieved. This module analyzes state transition data generated by actual actions and continuously updates model parameters, enabling the model to adapt to changes in equipment states. The module first retrieves the actual safety actions performed and their resulting state transition data from the data storage center. This data includes the system state before the action, the action parameters, the actual system state after the action, and timestamps. After acquiring the data, the module compares the actual state transition results with the previous predictions of the probabilistic world model, calculating the error value for each prediction. The prediction error is calculated using the mean squared error method, which involves squaring the difference between the predicted and actual values ​​for each state dimension and then averaging the results. Through this error calculation, the module can quantify the accuracy of the model's predictions and provide a basis for subsequent data selection.

[0083] During the data filtering phase, the module employs a priority selection mechanism based on prediction error. This mechanism calculates an importance score for each state transition sample, primarily based on the magnitude of the prediction error. Samples with larger prediction errors receive higher importance scores because they represent weak points in the model's current knowledge base. The module sets an error threshold, typically taking the top 20% of all sample error values ​​as the cutoff point. Only samples whose prediction errors exceed this threshold are marked as high-value samples. In addition to prediction error, the module also considers the sample's recentity, assigning higher weight to recently generated samples. Through this filtering mechanism, the module ensures that the training samples used for model updates are of high value, improving learning efficiency.

[0084] The incremental learning process employs streaming data processing. High-value samples selected are input into the probabilistic world model as a data stream. During model updates, a local parameter update strategy is used, adjusting only the model parameters relevant to these high-value samples, rather than retraining the entire model. Specifically, the system first determines the position of each high-value sample in the model's feature space, then identifies several kernel functions most relevant to that sample. During updates, only the parameters and weights of these relevant kernel functions are adjusted, while other parameters remain unchanged. This local update approach not only improves learning efficiency but also avoids destructive modifications to existing knowledge. The amount of data updated each time is controlled within 5% of the total training data to ensure the stability of model updates.

[0085] When new data encounters significant conflicts with existing model knowledge, the module initiates an adaptive model structure adjustment mechanism. Conflict detection is achieved by comparing the distribution characteristics of the new data with the distribution predicted by the original model. If multiple consecutive high-value samples exhibit patterns inconsistent with model predictions, and this inconsistency exceeds a preset conflict threshold, the system determines that model structure adjustment is necessary. Structure adjustment involves adding new kernel functions, whose parameters are initialized based on the characteristics of the conflicting data. The newly added kernel functions better characterize emerging state transition patterns, enabling the model to adapt to changes in equipment operating states. Simultaneously, the system periodically cleans up kernel functions with excessively low weights and small contributions to maintain model structure simplicity.

[0086] After the model update, the module reassesses the uncertainty distribution of the entire state-action space. The system iterates through all known state-action pairs, recalculating the prediction uncertainty for each pair using the updated probabilistic world model. The calculation process employs the same standards and methods as the initial training to ensure comparability of results. Based on the new uncertainty calculation results, the module dynamically adjusts the boundary division between the safe and unknown regions. This boundary adjustment uses a gradual strategy, allowing only small boundary movements in each update to avoid drastic boundary changes. The new boundary information is transmitted in real-time to the action safety simulation verification module as a basis for subsequent action safety judgments. Simultaneously, the historical record of boundary changes is saved for analyzing the changing trends of the system's safe region.

[0087] To ensure the quality of model updates, the module also establishes a robust validation and rollback mechanism. After each model update, the system tests the updated model using a separate validation dataset. The validation dataset contains real-world state transition data that has not been used in training within a recent period. Test metrics include prediction accuracy and uncertainty calibration. If the updated model performs below preset standards on these metrics, the system automatically rolls back to the previous model version. Simultaneously, the system records detailed logs for each update, including the update time, data used, parameter changes, and validation results. This log data is used to analyze the effectiveness of model updates and guide the optimization of subsequent update strategies.

[0088] The module also provides model performance monitoring, tracking the model's predictive performance in various state regions in real time. Monitoring metrics include the distribution of prediction errors, the degree of uncertainty calibration, and the model's adaptation speed to new data. When the model's predictive performance continuously declines in a specific region, the system proactively increases the sampling frequency in that region to acquire more training data. Simultaneously, the module periodically generates model performance reports, summarizing model improvements and existing problems over a period of time, providing decision support for continuous system optimization. Through this comprehensive online learning and optimization mechanism, the probabilistic world model can consistently maintain high predictive accuracy, ensuring the safe and reliable operation of the entire system.

[0089] In the intelligent decision-making strategy optimization module, the module is responsible for iteratively updating the agent's decision-making strategy based on the state transition results after the execution of safety actions, using a strategy optimization algorithm. This module first establishes an evaluation mechanism for state transition samples, quantitatively evaluating the results after each safety action. Evaluation indicators include two aspects: the degree of improvement in equipment health indicators and the magnitude of reduction in failure risk. The degree of improvement in health indicators is calculated by comparing the changes in key health parameters before and after the action execution, such as the percentage decrease in vibration amplitude and the degree of slowing down temperature rise. The magnitude of reduction in failure risk is measured by analyzing the marginal change in the system's distance from the failure threshold after the state transition. The system calculates a comprehensive score for each state transition sample, using a weighted average method, with health indicator improvement and risk reduction accounting for 60% and 40% of the weight, respectively. Based on the score results, the samples are divided into three priority levels: the top 20% are high-priority, the middle 60% are medium-priority, and the bottom 20% are low-priority. Samples of different priorities have different sampling weights in strategy updates, with high-priority samples having a weight five times that of low-priority samples.

[0090] During the policy network parameter update process, the module introduces safety constraints to ensure the safety and stability of the learning process. Safety constraints are implemented by limiting the action selection probability of the new policy within a safe region. Specifically, the system requires that the updated policy's action selection probability within a known safe region be no less than 0.95. This is achieved by adding constraints to the linear optimization objective, using the Lagrange multiplier method to transform the constrained optimization problem into an unconstrained optimization problem. The policy update employs a proximal policy optimization algorithm with a learning rate of 0.0001, using a mini-batch of 256 samples per update. During the update process, the system monitors the constraint satisfaction in real time. If a constraint violation is detected, the optimization parameters are automatically adjusted until the requirements are met. Furthermore, the system records the constraint satisfaction status during the policy update process to analyze the stability of the policy learning process.

[0091] The policy distillation and transfer mechanism periodically transfers knowledge from a well-trained main policy network to a lightweight policy network. The distillation process is executed every 24 hours, using decision data generated by the main policy network over the past 24 hours as the training set. The lightweight network employs a depthwise separable convolutional structure, reducing the number of parameters to 20% of the main network while maintaining the same input and output dimensions. During training, knowledge distillation techniques are used, employing the output distribution of the main network as soft labels, combined with real-world state-action pairs for co-training. After training, the lightweight network is deployed to edge computing devices for real-time decision inference. Simultaneously, the system compares the decision consistency between the lightweight network and the main network to ensure the quality of knowledge transfer. When the consistency falls below 0.9, the distillation process is restarted.

[0092] The validation and rollback mechanism for the new strategy ensures that the system maintains reliable decision-making performance at all times. After each strategy update, the system performs a comprehensive performance test. This test uses an independent test set containing 1000 recently collected state transition samples. Test metrics include three dimensions: decision accuracy, safety metrics, and efficiency metrics. Decision accuracy is measured by comparing the consistency between the strategy output and expert decisions; safety metrics assess the conservatism of the strategy under dangerous conditions; and efficiency metrics measure the optimization performance of the strategy under normal conditions. If the performance of the new strategy degrades by more than 5% in any dimension, or if a potential security risk is detected, the system automatically triggers the rollback mechanism. The rollback process restores the strategy parameters to the previous stable version and records the failure information of this update for analyzing the reasons for failure and improving the update algorithm. The system retains the strategy parameters from the most recent 10 versions to ensure sufficient rollback options.

[0093] It should be noted that this system monitors equipment status in real time through a multi-source data acquisition and fusion module. When an abnormal pattern appears in the system state sequence (such as a sudden increase in vibration amplitude or a continuous rise in temperature), the probabilistic world model predicts the uncertainty of state transitions and identifies high-risk actions through safe zone delineation. The action safety simulation verification module simulates the execution path of candidate actions and calculates the cumulative uncertainty. If the uncertainty exceeds a threshold, the system blocks the action and triggers a safety strategy, generating alternative actions to avoid failure. Simultaneously, the online model learning and optimization module dynamically updates the model based on actual data, expands the safe zone, and achieves early warning and autonomous diagnosis of faults.

[0094] The working principle of this invention is as follows: A multi-source data acquisition and fusion module collects and processes sensor data from various subsystems of the tunnel boring machine in real time, constructing a system state sequence containing equipment operating parameters and health indicators. Based on this sequence and historical action sequences, a probabilistic world model construction module uses a non-parametric method to establish a probabilistic world model capable of predicting state transition probability distributions and outputting uncertainty quantification values. A dynamic safety zone partitioning module divides the state action space into safe and unknown zones based on the uncertainty quantification values. An action safety simulation verification module calculates the cumulative uncertainty of candidate actions through multi-step rolling prediction. A safe action decision and execution module determines whether to execute or block an action based on an uncertainty threshold. A model online learning and optimization module dynamically updates model parameters through data filtering and incremental learning. An intelligent decision strategy optimization module iteratively updates the decision strategy based on a safety constraint strategy optimization algorithm, ultimately achieving continuous optimization of fault prevention and autonomous diagnosis functions.

[0095] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A fault prevention and diagnosis control system for mining shield machines based on Kuanghong Intelligent, characterized in that, include: The multi-source data acquisition and fusion module collects multi-source heterogeneous sensor data in real time during the operation of the tunnel boring machine through the Kuanghong operating system, and constructs a system status sequence that includes equipment operating parameters and health indicators. The probabilistic world model building module establishes a probabilistic world model based on the system state sequence and historical action sequence, which can predict the state transition probability distribution and output the quantified value of the prediction uncertainty. The safe zone dynamic division module divides the state action space into a high-confidence safe zone and a low-confidence unknown zone based on the predicted uncertainty quantification value. The action safety simulation verification module simulates the execution of candidate actions through the probabilistic world model when the agent generates candidate actions, and calculates the cumulative uncertainty on the predicted state path. The safety action decision and execution module compares the accumulated uncertainty with a preset safety threshold: if it is lower than the threshold, the candidate action is output to the control system as a safe and executable action; if it exceeds the threshold, the candidate action is blocked and a preset safety strategy is triggered to generate an alternative action. The online learning and optimization module feeds back the actual safety actions performed and the resulting state transition data to the probabilistic world model, dynamically updating the model parameters and expanding the boundaries of the safety region. The intelligent decision-making strategy optimization module uses a strategy optimization algorithm to iteratively update the agent's decision-making strategy based on the state transition results after the execution of safety actions, thereby continuously optimizing the fault prevention and autonomous diagnosis functions.

2. The fault prevention and diagnosis control system for mining shield machines based on Kuanghong Intelligent as described in claim 1, characterized in that, The process of obtaining the system state sequence is as follows: Using the soft bus protocol of the Kuanghong operating system, the raw measurement values ​​of vibration, temperature, pressure and displacement sensors deployed in the main drive system, propulsion system, cutterhead system, hydraulic system and lubrication system of the tunnel boring machine are collected in parallel, and the data of each system are marked with a unified time stamp. The raw measurement values ​​are preprocessed in real time to eliminate measurement noise caused by downhole electromagnetic interference and mechanical shock. At the same time, the effective value, peak factor and kurtosis index of each sensor are calculated by feature extraction to form a set of equipment health indicators. The preprocessed sensor measurements and the set of equipment health indicators are combined in time sequence to construct a multi-dimensional state vector containing instantaneous operating parameters and medium- to long-term health trends. Each dimension of the state vector is normalized. The multidimensional state vector is continuously updated with a fixed sampling period to form a system state sequence that can reflect the complete operating status of the device in real time, and the system state sequence is transmitted to the data storage center.

3. The fault prevention and diagnosis control system for mining shield machines based on Kuanghong Intelligent as described in claim 1, characterized in that, The construction process of the probabilistic world model is as follows: The multidimensional state vectors in the system state sequence are combined with the historically executed action sequences according to the time correspondence to form a state-action pair sequence, which is used as the model training input. A state transition probability model is established using a non-parametric method based on kernel density estimation. By calculating the similarity weights of state-action pairs in the feature space, the probability distribution of the next state under a given state-action condition is predicted. The uncertainty of the probabilistic world model for each state-action pair is quantified by calculating the degree of dispersion of adjacent sample points in the predicted state distribution. The quantified uncertainty value is represented by a normalized value between zero and one. The probabilistic world model is updated online using state transition data within a recent time period, enabling it to adapt to changes in device state over time.

4. The fault prevention and diagnosis control system for mining shield machines based on Kuanghong Intelligent as described in claim 1, characterized in that, The division of the state-action space into a high-confidence safe region and a low-confidence unknown region specifically includes: Determine whether the quantified value of the prediction uncertainty is greater than or equal to a preset threshold. If yes, it is recorded as a high-confidence safe region; otherwise, it is recorded as a low-confidence unknown region. The preset threshold is determined through historical data statistics or experimental calibration, and its specific value is within the range of 0.6 to 0.

8.

5. The fault prevention and diagnosis control system for mining shield machines based on Kuanghong Intelligent as described in claim 1, characterized in that, The step of simulating the execution of candidate actions through the probabilistic world model and calculating the cumulative uncertainty on the predicted state path specifically includes: Starting from the current state, the candidate actions are input into the probabilistic world model to obtain the next state prediction, and the next state prediction is used as a new starting point to iteratively execute multi-step state transition prediction, forming a multi-step predicted state path. In each prediction step, the state prediction uncertainty quantification value is recorded, and the cumulative uncertainty on the entire prediction path is calculated using a time decay weighted algorithm. The uncertainty of the recent prediction steps is assigned a weight of 0.9 to 1.0, and the uncertainty of the long-term prediction steps is assigned a weight of 0.1 to 0.

4. When a critical state node is predicted, multiple possible action branches are generated to continue prediction. The cumulative uncertainty of each branch path is calculated and the maximum value is taken as the final cumulative uncertainty of the corresponding candidate action. The multi-step prediction path and the uncertainty quantification values ​​of each node are displayed in the form of time-series animation, providing an intuitive security verification process.

6. The fault prevention and diagnosis control system for mining shield machines based on Kuanghong Intelligent as described in claim 5, characterized in that, The optimization and adjustment function for the candidate actions specifically includes: When the accumulated uncertainty of the candidate action exceeds the safety threshold, the action fine-tuning algorithm is activated to generate multiple alternative actions with small variations in the neighborhood of the candidate action. Each alternative action is sequentially input into the probabilistic world model for multi-step prediction, and the cumulative uncertainty corresponding to each alternative action is calculated. The alternative action with the lowest cumulative uncertainty and below the safety threshold is selected as the optimized safety action. If the cumulative uncertainty of all alternative actions exceeds the threshold, the preset safety policy is activated. Record the process data for each action optimization and adjustment, including the original action, alternative actions, and their uncertainties, to form an action optimization knowledge base.

7. The fault prevention and diagnosis control system for mining shield machines based on Kuanghong Intelligent as described in claim 1, characterized in that, The step of feeding back the actual security actions and the resulting state transition data to the probabilistic world model, dynamically updating the model parameters and expanding the security region boundary, specifically includes: By comparing the actual state transitions with the model's predicted states, the state transition data of the top 20% percentiles with prediction errors greater than all sample error values ​​are selected as high-value samples. The high-value samples obtained from the screening are input into the probabilistic world model as streaming data, and only the local parameters of the model associated with these samples are updated in a targeted manner. When new data conflicts with existing model knowledge, the model structure adaptive adjustment process is initiated, and new kernel functions are added to adapt to changes in state transition rules. The prediction uncertainty of all known state-action pairs is recalculated based on the updated probabilistic world model, and the boundary division between the safe zone and the unknown zone is dynamically adjusted.

8. The fault prevention and diagnosis control system for mining shield machines based on Kuanghong Intelligent as described in claim 1, characterized in that, The continuous optimization of the fault prevention and autonomous diagnosis functions specifically includes: Based on the degree of improvement in equipment health indicators and the extent of reduction in failure risk in the state transition results, different priority weights are assigned to each state transition sample. When updating policy network parameters, introduce security constraints to ensure that the probability of the new policy's action selection within the safe zone is not lower than a set threshold. Implement periodic policy distillation and transfer to distill the knowledge of a well-trained master policy network into a lightweight policy network, enabling rapid deployment and real-time decision-making; Regularly test the effectiveness of new policies in real-world environments, and automatically roll back to the previous stable version when performance degradation or security risks are detected.

Images